example of openBLAS speeding up models?

[James-Thorson-NOAA]

Hi all,

I can't seem to get an improvement in sdmTMB (or other TMB models) using the openBLAS instructions that are very thoroughly and helpfully provided. Below, I show that the matrix-multiplication is faster (as expected) but Pcod spatial model is slower using the updated files relative to base versions.

Does anyone have additional experience to share regarding enabling openBLAS on a windows machine using NOAA IT privileges?

Using the standard Rblas.dll and Rlapack.dll files:

and using the updated versions:

[seananderson]

Have you ruled out that openBLAS is using all available cores? That can slow things down. There's a bit at the bottom of the Gist about that:

You may find that openBLAS uses all available cores, which may not be computationally efficient with TMB. If so, you can specify BLAS should only use 1 core with the RhpcBLASctl package:

RhpcBLASctl::blas_set_num_threads(1)
RhpcBLASctl::omp_set_num_threads(1)

You can either set that at the beginning of each script or (better yet) place it in your .Rprofile using:

usethis::edit_r_profile()

Assuming the above is the default behaviour, maybe we should make those instructions more prominent or try to detect that for the user.

Apple's vecLib seems to default to 1 core.

Also, my memory is the speed up is more noticeable with an AR1 or random walk spatiotemporal field, but at the very least it shouldn't be slower for a simple spatial model!

[James-Thorson-NOAA]

Below is with the spatio-temporal RW turned on. I've tried:

• playing with RhpcBLASctl::blas_set_num_threads(24) and RhpcBLASctl::omp_set_num_threads(24) vs. RhpcBLASctl::blas_set_num_threads(1) and RhpcBLASctl::omp_set_num_threads(1) and don't see a big change in CPU use in the "Task manager" ... I don't know if that's a sign that something is wrong?
• Re-installation sdmTMB locally in case I need to recompile the CPP for the changes to take effect
• Setting OPENBLAS_NUM_THREADS as environment variable as per David-Cortes advice
  ... and none of these have seemed to fix things.

Of course it's not a huge issue (I can always use MRAN locally), but I'd be happy to hear if there's other intermediate diagnostics that might illustrate where the issue is arising.

With updated BLAS:

and default BLAS:

[seananderson]

Hmm... I'm not sure. I'll have to try this myself on a DFO Windows laptop.